Generative Engine Optimization: What Actually Moves the Needle, and What's Snake Oil

The difference between ranking on a page and getting named in an answer

Search used to end with a list. You typed a query, ten blue links came back, and the game was to sit as high in that list as possible. A written AI answer works differently. Someone asks ChatGPT or Gemini or Perplexity "what's a good Shopify app for chargeback recovery", and the model writes a paragraph that names two or three products. Everyone else is invisible. There is no page two.

That is the whole problem generative engine optimization is trying to solve: not "how do I rank", but "how do I get named in the answer at all". It goes by a few names — generative engine optimization, answer engine optimization, sometimes just AI search visibility — and the market around it is filling up fast with people promising things nobody can actually deliver. This post is an attempt to separate what demonstrably influences whether you get cited from what is being sold as certainty and isn't.

The stakes are not abstract. McKinsey projects roughly $750B in US revenue will flow through AI-powered search by 2028, while only about 16% of brands track how they appear in AI answers. If that's even close to right, a lot of demand is going to route through a channel most companies aren't watching.

What GEO actually is

Getting cited in a generated answer is closer to earning a reference than winning a ranking. The model isn't sorting a database of pages by a score you can climb. It's producing text from what it has learned about the world, then — in the case of the search-connected assistants — checking a handful of live sources and deciding which ones are worth naming.

So the unit of success is a mention or a citation inside prose, and the field of play is tiny. A blue-link results page has room for ten, twenty, fifty entries. A written answer has room for maybe three brands before the reader's eyes glaze. Answer engine optimization is the discipline of being one of those three, which is a harder and more selective target than being on page one ever was.

It's worth being blunt about the mechanism up front, because most of the snake oil grows in the gap where people don't understand it:

• No one controls what a model says. Not you, not the vendor you hire, not the people who built the model. Output is probabilistic and it shifts between versions. Anyone who tells you otherwise is describing a product that doesn't exist.
• There is no "submit to ChatGPT" button. You cannot register your brand with an assistant the way you once submitted a sitemap to Google. There is no console, no queue, no intake form. The influence is all indirect — you change what the web says about you, and hope the model learns from it.
• Guaranteed placement is not a real thing. A specific spot in a specific answer cannot be promised, because the thing making the decision is outside everyone's control. Guarantees in this space are marketing, not engineering.

Hold onto those three facts. Everything below is either consistent with them or it's a scam.

What demonstrably influences whether you get cited

Nobody can guarantee a citation, but that doesn't mean it's random. There are levers that measurably change the odds, and they're mostly unglamorous. The through-line is simple: assistants tend to name brands they can describe confidently and confirm from sources they already trust.

• Consistent, machine-readable facts about your brand across the web. If your product's name, category, pricing model, and core claim read the same on your own site, your app-store listing, your reviews, and third-party write-ups, a model can state them without hedging. If those facts contradict each other across the web, the model gets uncertain — and an uncertain model reaches for a competitor it can describe cleanly.
• Structured data and schema. Marking up your pages with schema.org types — Product, Organization, FAQ, Review — turns prose a crawler has to interpret into facts it can read directly. This is the same structured data that helped with traditional rich results; it's plumbing, and it's within your control, which makes it one of the few honest starting points.
• Entity clarity. The model needs to know, without ambiguity, who you are and what you sell. "Wizovia builds and runs its own Shopify software and does AI search visibility work" is an entity a model can hold. A vague "we help brands grow" is not. Clear comparison pages, factual product pages, and a plainly stated category do more here than any clever phrasing.
• Citations from sources the assistants already trust. Being referenced by reputable third-party pages, independent reviews, and — where it's genuinely warranted — Wikipedia or Wikidata gives a model corroboration it will lean on. This is earned, not bought. A legitimate Wikidata entry for a company that actually meets the notability bar is fair; manufacturing one for a company that doesn't is the kind of thing that gets reversed and can hurt you.
• Genuine reputation. Real reviews, real coverage, real usage. Models are trained on a web where those signals are diffuse and hard to fake at scale, so the durable way to be described as credible is to be credible. There's no shortcut here that survives contact with a model retrained six months later.

None of these is a switch. They are conditions that make a citation more likely, and they compound. That's the honest framing — "demonstrably influences", not "guarantees" — and it's the framing we hold ourselves to on our AI search visibility work.

What is snake oil, and why it falls apart

The tells are consistent once you know the mechanism. Each of these pitches assumes a control that no one has.

• "Guaranteed placement in AI answers." Impossible, for the reason stated twice already: the model deciding is outside anyone's control, and its output moves between versions. A guarantee here is a promise to control something the promiser cannot touch.
• "We'll submit you to ChatGPT." There is nothing to submit to. No intake exists. This pitch is selling a mechanism that isn't there — usually to people who assume AI search works like the search consoles they already know.
• Keyword-stuffed pages "written for the LLM." Pages packed with repeated phrases to appeal to a model tend to read as low-quality to the exact systems they're meant to game, and they degrade the trust signals that actually help. You end up looking less credible to the thing you were trying to impress.
• Buying mentions. Paid placements dressed up as organic references are fragile. When the model retrains, or the source gets discounted as low-trust, the mention evaporates and you've paid for something with no durability. You're renting a signal, not building one.
• Anyone promising a specific ranking. "We'll get you to the top of AI results" imports a mental model — a rankable list — that doesn't describe how a generated answer is produced. If a vendor is promising a rank, they either don't understand the medium or are counting on you not to.

The common thread: every one of these sells certainty over a process that is inherently uncertain. The moment a pitch promises a guaranteed outcome, it's contradicting how the technology works.

How to evaluate a GEO vendor honestly

You can't judge this work by promises, because the honest version of it doesn't make many. Judge it by measurement and candor instead. Three questions do most of the filtering.

• Do they measure? A serious vendor tracks how you actually appear across assistants — which queries name you, which name competitors, how that changes over time. If they can't tell you your current AI search visibility before proposing to improve it, they're guessing. Ask what their baseline is and how they capture it.
• Do they show movement, not placement? Because no single answer can be promised, the honest signal of progress is a trend — mentions appearing where there were none, over weeks, across a set of tracked prompts. Movement in a measured baseline is real. A screenshot of one good answer proves nothing; answers vary by session and re-roll.
• Do they admit what they can't control? This is the strongest tell. A vendor who says plainly "we can't guarantee a citation, here's what we can influence and here's what's outside anyone's hands" understands the medium. A vendor who guarantees results is either uninformed or willing to mislead you. In a field this new, the willingness to name the limits is the credential.

Generative engine optimization is real work with real levers — consistent facts, structured data, entity clarity, earned trust, genuine reputation. It is also a magnet for confident promises that the underlying technology cannot support. The way through is to keep both facts in view at once: this demonstrably influences whether you get named, and no one can promise you will be. A vendor comfortable saying both is one worth talking to. If you want to see what that looks like measured against your own brand, that's the AI search visibility work.